IP Address Allocation

Facing continuous and intensified discussions on IP address allocation
issues, IETF Chair Paul Mockapetris called this session to summarize the
current views and activities around the subject.

Peter Ford first described CIDR, Classless InterDomain Routing, as a
solution to the current crisis of routing table overflow.  In addition
to being classless, CIDR also requires topology-based IP address
allocation in order to achieve large aggregation, hence routing table
size reduction.  As a result, one must renumber as topologies change
(e.g., change of providers).  Ford believes that most customers do not
care about permanent IP addresses; the limiting factor is merely lack of
adequate renumbering tools.

Elise Gerich then presented a summary of RFC 1814, ``Unique Addresses
Are Good''.  She listed three attributes of an IP address that have
traditionally been considered desirable:  uniqueness, portability, and
routability.  With the growth of the Internet, it has become
increasingly difficult to satisfy all three of these desired attributes.
Routability requires address uniqueness, however, in reality acquiring a
unique address does not guarantee that a service provider will route the
address.  Portable addresses make address aggregation difficult or
impossible, but organizations are demanding portable addresses to avoid
the pain and necessity of renumbering.  The tensions between the desire
for address ownership and the need to compress routing table size will
continue until technology advances lead to either trivial renumbering or
support for continuation of flat routing.  Until at least one
technological solution exists, we will continue to make difficult
compromises concerning the attributes of an IP address.

As the final speaker of the session, Lixia Zhang gave a brief interim
progress report of a study group she had organized under the instruction
of the IAB. The group's mission is to evaluate each of the proposed
solutions on the table, mainly provider-based address assignment (CIDR),
dynamic provider-based address binding, and metro-based address
assignment.  She pointed out that up to now, the IP address has served
as an invariant, unique identification for the end host.  TCP design
makes use of this assumption, so do many other protocols and
applications.  As a result, nobody today has a complete list of all the
possible places in the protocol architecture that have the IP address
hardwired or embedded in it.  Therefore, contradicting Peter's
assumption that most customers do not care about permanent IP addresses,
dynamically changing addresses, as required by provider-based
assignment, changes the architecture we used to know and causes serious
problems at the user ends.  In addition to the need for
autoconfiguration on tools at low levels, renumbering requires changes
to high-level protocols.  It also puts further reliance on the DNS
system to keep up-to-date address binding.  To avoid circular
dependency, DNS servers themselves will require special treatment, such
as provider-independent addresses, assured connectivity, issues that are
yet to be explored.  As part of this interim report, Steve Deering gave
a briefing on the main features of metro-based approach.  Its foremost
advantage is avoidance of renumbering, hence avoiding all the problems
mentioned above.  The main drawback/difficulty is the required
connectivity to the MIX at each metro area where a provider has a
presence.  The study group's next step includes documentation of each of
the proposals and their open issues; the final report is expected by the
December IETF meeting.